Persisted Agent Chats
We just shipped persisted workspace chats for Dreambase agents. On paper it
sounds like basic chat history. Under the hood it is a meaningfully different
architecture, and it directly improves dashboard generation quality. Here is
what changed and why it matters.
The Problem
Every agent conversation in Dreambase used to be ephemeral. Close the tab, lose
the context. Each session started cold: no memory of prior questions, no memory
of how you described your schema or business logic, no memory of conclusions
already reached.
For a product where the core value is an agent reasoning over your Supabase
Postgres data plus connected APIs and MCP sources, that is a real limitation.
Context compounds. Stateless sessions cannot compound.
What Changed
Workspace chats are now persisted to the database, but the persistence layer is
not a simple message log. It carries the same context management, retrieval,
compression, agent framework, and tool access that already powers our dashboard
chat experience, which received its own upgrade alongside this release.
Practically, this means:
- Prior conversation history is retrievable and used as active context on future turns, not just displayed as a transcript.
- Long conversation histories are compressed rather than truncated, so context windows stay efficient without losing the substance of what was discussed.
- The agent framework and tool access available in a persisted chat matches what is available in our dashboard planning chat, so there is no capability gap between a quick question and a full planning session.
- The system is designed to self-improve: retrieval and compression get better as the underlying models and techniques improve, without requiring you to change how you use chat.
Why This Improves Dashboard Generation Specifically
Plan Mode, our guided dashboard planning flow, is built entirely on chat. Before
this release, every Plan Mode conversation started from a blank context. Now it
inherits everything the agent has learned from prior conversations in that
workspace: your metric definitions, your data quirks, the questions you have
already asked and answered.
The effect cascades. Better context means sharper clarifying questions during
planning. Sharper planning means the generated dashboard matches intent on the
first pass more often. Fewer generation-refinement loops means less time between
asking for a dashboard and having one you trust.
This is the kind of improvement that will not show up as a single flashy
changelog line. It shows up as a measurable drop in edit rounds per dashboard
and a measurable increase in first-pass accuracy across our internal evals.
Why We Built It This Way First
There is a reason this needed to be solid internally before we did anything else
with it.
We are building a Dreambase MCP server and public REST API that will expose our
agents, dashboards, and datasets externally. The goal is for you to set up,
configure, and call Dreambase directly from the AI apps, workflows, schedules,
routines, and automation loops you already run, the same way Dreambase's own
agents currently call PostHog, Linear, or Stripe through MCP.
That only works if the underlying chat and agent architecture is durable,
context-rich, and reliable under real usage. Persisted workspace chat is that
architecture, proven internally first. It is the direct predecessor to the MCP
server and API work coming next, not a separate feature.
What's Next
The MCP server and public API are next up. If you want early access or want to
integrate Dreambase agents and datasets into your own tooling ahead of general
availability, reach out.
or
Read the full story:
https://dreambase.com/blog/improved-persisted-agent-chats
Tags: ai, agents, supabase, analytics, mcp, llm
Top comments (0)